augustus odena

312 posts

augustus odena banner
augustus odena

augustus odena

@gstsdn

Something new. Previously: AI research at TBD Labs / Meta; cofounder at @AdeptAILabs; Invented Scratchpad / Chain-of-Thought; Google Brain

Katılım Ekim 2015
3.3K Takip Edilen12.3K Takipçiler
Sabitlenmiş Tweet
augustus odena
augustus odena@gstsdn·
Yesterday I resigned from TBD Labs / @metaai. I wasn't there for very long, but I think I got a few useful things done! It's an impressive group of people and it's especially impressive that it got assembled as quickly as it did with such a high talent bar at a large company. "Founder Mode" is real and good. I will certainly miss my coworkers there. I think now is an unusually high-leverage time to pursue ambitious new projects at the intersection of AI and other technologies. Please reach out to me if you're interested in that sort of thing, and I expect I will have something more detailed to share in not-too-long.
English
26
6
632
283.1K
augustus odena retweetledi
hope hopes hoping
hope hopes hoping@hopes_revenge·
personally i am not really impressed when a computer is good at math — since a computer is math . i am a lot more intrigued by things like , for example , a rat being an exceptional cook — since that’s not really expected
English
97
1.8K
31.4K
505.5K
augustus odena
augustus odena@gstsdn·
the Notion agent is quite bad in my experience. This is interesting because it suggests some combination of: - Even a basic agent harness is actually quite hard to get right - grafting on an agent harness to software that wasn't designed "agent-first" is too hard?
English
0
0
3
880
augustus odena
augustus odena@gstsdn·
As far as technical artifacts go, a "frontier" LLM is just not that hard to produce, relatively speaking. There are more "frontier labs" than companies capable of building a leading edge chip fab, or a reusable launch vehicle, or a large commercial turbofan, etc. I think this has some pretty obvious economic implications...
English
5
1
41
5.2K
augustus odena retweetledi
Alex Imas
Alex Imas@alexolegimas·
AGI will very likely be a commodity. This is not a value judgement, it's just the outcome of an economically-valuable product + competitive markets. Implicitly people see this coming and are leaning into it, hence the big push from all the labs to go vertical. A few things can prevent this: 1) regulation or 2) a moat, so it's definitely not a certainty. But absent that, this is where things are going.
Claude@claudeai

Beginning July 20, Claude Fable 5 will be included in all Max and Team Premium plans, at 50% of limits. Pro and Team Standard users will continue to have access to Fable via usage credits, and will receive a one-time $100 credit. Demand for Fable has been challenging to predict, which is why we rolled it out to subscription plans in stages, extending access several times as we secured additional capacity.

English
59
68
644
200.5K
augustus odena retweetledi
Daniel Keller
Daniel Keller@dnlklr·
Is nothing sacred to these AI freaks???? Mark Zuckerberg and Larry Ellison just built this monstrous data center in the very heart of Paris—only steps away from Notre-Dame and the Louvre. Can’t the EU do something?
Daniel Keller tweet media
English
300
434
6.1K
803.9K
augustus odena
augustus odena@gstsdn·
This is really really cool. It seems plausible to me that devices like these largely replace phones in the next ~5 years and that they become one of the main ways we interact w/ LLMs (because they're low friction to talk to + they have visual context)
Thomas Suarez@tomthecarrot

We’re sharing an early preview of Raven Prism, our stylish ambient computer in glasses form, ahead of its launch later this year. Raven Prism is Linux-based with a ARM64 architecture and devs have already built apps for it. If you’re at AWE, come by our booth for a demo! raven.computer @raven_computer

English
2
1
5
1.6K
augustus odena retweetledi
WIRED
WIRED@WIRED·
With $500 million in funding and a reported $2.5 billion valuation, Flourish wants to reinvent AI by putting real neurons under the microscope. wired.com/story/jeff-bez…
English
8
26
104
71.2K
augustus odena retweetledi
Tri Dao
Tri Dao@tri_dao·
After some mathematical rewrite, turns out all of transformer is a series of gemm + epilogue. Given a few optimized primitives, LLMs (and novice humans) can write speed-of-light kernels for all transformer ops!
Han Guo@HanGuo97

LLM training is built on fast MatMuls. But many surrounding ops still run as memory-bound kernels. CODA reparameterizes them to hide in the matmul’s shadow, fused into its epilogue before results leave the chip. Bonus: LLMs can write fast CODA kernels too (approaching SoLs).

English
17
128
1.2K
134.3K
augustus odena retweetledi
Arram
Arram@arram·
Asked Claude: 'There's a meme called the "fix everything easily switch". What policies do you think are the best candidates for being a real fix everything switch in the US? Give me your top ten, your confidence, your reasoning, and why a given policy has not been implemented.'
Arram tweet media
English
144
281
2.3K
936.8K
augustus odena
augustus odena@gstsdn·
Maybe different types of intelligence are close to a Pareto frontier / have hard-to-avoid tradeoffs. For example, even though human brains are much better at continual learning than LLMs, they are extremely hard to duplicate. Could this tradeoff be unavoidable? Maybe the tradeoff is real, but you can route around it by spending 5 more OOMs of compute to run a high-fidelity brain simulation which can be copied. But then you might hit another problem: does the way continual learning works preclude you from merging the copies together once they've been deployed in separate environments?
English
1
0
4
909
augustus odena
augustus odena@gstsdn·
Why don't we see bigger productivity gains from LLM usage? Two issues, IMO. 1. To really get leverage out of delegating to someone (e.g. a vendor or an employee), you have to give up on doing a detailed check of the work every time. But you mostly can't do this with LLMs because they make too many mistakes and (for many types of work) because of social constraints. The latter will get better for sure. The former will get better for things you can easily RL... 2. Because of issues with continual learning, it's hard to give feedback to keep the delegation relationship on track. We have ways to hack around this - e.g. cursor rules - but they work much less well than giving feedback to a human
English
1
0
10
924
augustus odena
augustus odena@gstsdn·
Incredible stuff coming into SFO last night.
augustus odena tweet media
English
3
0
14
747
augustus odena retweetledi
Amjad Masad
Amjad Masad@amasad·
“Mum, can we have the SpaceX IDE?” “No we have a space IDE at home” Space IDE at home:
Amjad Masad tweet media
English
8
4
153
33.1K
augustus odena retweetledi
roon
roon@tszzl·
@Noahpinion everyone underestimates the self commodifying nature of ai technology
English
39
30
613
31.3K
augustus odena
augustus odena@gstsdn·
We need to put a canal in SF
augustus odena tweet media
English
1
0
9
1.1K
augustus odena
augustus odena@gstsdn·
@lishali88 Ehhh the jet lag is pretty brutal for her I’m not sure I’d recommend it vs like Mexico City or something (which we also did somewhat recently with her)
English
1
0
1
101
Lisha
Lisha@lishali88·
@gstsdn How is it traveling with a toddler? Was planning to bring mine there
English
1
0
1
177

Keşfet

@Noahpinion @lishali88 @elonmusk @BarackObama @taylorswift13 @cristiano @BillGates @NASA